A Data-Driven Framework for Classification and Retrieval of Inorganic Materials Synthesis Recipes

Abstract

Large text-mined collections of materials synthesis procedures create opportunities for data-driven analysis, but a recipe can be represented through lexical text, structured chemical entities, numerical process conditions, or domain-language embeddings. This paper presents a leakage-controlled framework for comparing these representations in classification and nearestneighbor retrieval. Experiments use 35,675 solution-based synthesis recipes and 19,488 solid-state recipes. Source-paper identifiers are canonicalized before five-fold grouped cross-validation and are also used to exclude same-paper neighbors during retrieval. Six independent representations are evaluated: TF-IDF, hashed categorical features, numerical process features, raw MatSciBERT, unsupervised MatSciBERT-SimCSE, and generic MiniLM. For hydrothermal-versus-precipitation classification, TF-IDF with logistic regression achieves the highest macro-F1 (0.8429) and ROC-AUC (0.9223), while raw MatSciBERT reaches 0.8188 macro-F1 and exceeds generic MiniLM. In retrieval, MatSciBERT-SimCSE gives the strongest sustained synthesistype precision at ranks 5 and 10 (0.7300 and 0.7207), whereas lexical, categorical, and numerical representations specialize in chemistry or process similarity. SimCSE also reduces the mean random-pair cosine similarity of solution-recipe embeddings from 0.9430 to 0.0813. The results show that no single representation is universally optimal and that synthesis-recipe similarity should be evaluated with task-specific criteria and strict article-level leakage controls. Reliable retrieval of comparable procedures can subsequently support sustainability-oriented comparison of synthesis conditions and precursor choices.